Purpose: To identify the needs and opportunities in standardizing and improving our external beam radiotherapy (EBRT) treatment planning procedure across all 7 campuses of our institution, a knowledge-based anomaly detection algorithm to provide decision support for our manual retrospective plan auditing process was developed and implemented.
Methods: 710 lung SBRT (Stereotactic Body Radiation Therapy) plans that were delivered from 2015 to 2020 have been acquired automatically from our clinical treatment planning systems. A total of 44 plan parameters were extracted from each plan and pre-processed. A knowledge-based anomaly detection algorithm, namely “isolation forest” (iForest), was then applied. Each plan was recursively isolated by randomly selecting a plan parameter and a split value, represented by an isolation tree (iTree). Subsequently, an iForest is assembled from a collection of developed iTrees and an anomaly score is determined from the path length between the root node and a terminating node on the iTree. Anomalous plans are more susceptible to such isolation mechanism, thereby having higher anomaly scores. To evaluate the feasibility of the proposed method, top 20 plans ranked with the highest anomaly scores together with the extracted plan parameters were validated manually by two plan auditors.
Results: The two auditors certified that 12 out of 20 plans with the highest iForest anomaly scores have similar “alarming” qualities to further investigate and improve our planning procedures and staff training. A manual audit of one chart is estimated to take 45 minutes, while with the guidance of iForest it is estimated to save approximately 7.5 minutes per chart, a time savings of 16.7%. As our internal audit process typically consists of review of 250 charts annually, the total time savings are substantial.
Conclusion: The iForest method effectively detects anomalous plans with high sensitivity and provides decision support to our cross-campus manual plan auditing procedure.